NVidia shows how augmenting real data with synthetic data outperforms the real data alone.
Earlier domain randomization paper from the same group at nVidia that shows synthetic data + real data beats real data.
Use of synthetic data to improve generalization for rare classes:
Synthetic point cloud generation using GTA:
Meta-Sim: Learning to Generate Synthetic Datasets:
Virtual Worlds as Proxy for Multi-Object Tracking Analysis (VKITTI paper):
Playing for Data: Ground Truth from Computer Games:
Learning To Simulate:
Domain randomization for transferring deep neural networks from simulation to the real world:
Synscapes: A photorealistic synthetic dataset for street scene parsing (7DLabs papers):
Building Generalizable Agents with a Realistic and Rich 3D Environment:
The SYNTHIA Dataset: A Large Collection of Synthetic Images for Semantic Segmentation of Urban Scenes:
SceneNet: Understanding real world indoor scenes with synthetic data:
A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation:
Driving in the matrix: Can virtual worlds replace human-generated annotations for real world tasks?: